This paper focuses on the approach to attitude estimation. When describing the attitude of a dynamic object, the quaternion has a better numerical continuity and stability compared with other conventional Euler angles’ methods. However, acceleration observation vector has a nonlinear relation to attitude quaternion and as a result, the representative linear methods e.g. Kalman filter (KF) and complementary filter (CF) are no longer applicable. In addition, the disturbances of accelerometers and magnetometers also greatly degrade the attitude estimation reliability, leading to solution biased even divergence. In this contribution, the general heterogeneous MARG data fusion strategy is proposed, to minimize the noises influences of nonlinear system imposing on the attitude estimation of MARG sensors. To overcome the nonlinear estimation problem, the unscented Kalman filter (UKF) for attitude determination is proposed based on the unscented transformation. Furthermore, a real-time disturbance detection rules are established for the external acceleration and magnetic field distortion. Finally, the real experiments are carried out to evaluate performances of our proposed attitude estimation method.